1. Editorial Note:
The following article by Susan Hurley, “The Shared Circuits Model: How Control, Mirroring, and Simulation Can Enable
Imitation, Deliberation, and Mindreading,” with its commentaries and response was produced under unusual and sad
circumstances. Susan Hurley passed away in August 2007 following a long struggle with cancer after her target article
had been completed, and the list of those invited to comment had been assembled. Because she had foreseen the need for
help in producing her response to the commentaries, she enlisted Andy Clark, Professor at Edinburgh University, for
this purpose, with BBS’s full encouragement. Julian Kiverstein, another colleague at the University of Edinburgh with par-
ticular interest in the shared circuits model, volunteered to help as well in the composition of the response to commentators.
Commentators were specifically enjoined from writing eulogies and asked to produce the lively intellectual dialogue that
Susan Hurley certainly had sought in sending her work to BBS. Kiverstein and Clark undertook not to emulate a response
from Susan Hurley, but rather to clarify misunderstandings, organize the commentaries thematically, and show where the
research might lead. We are grateful to all commentators, and particularly Kiverstein and Clark, for their graceful
execution of what even in the normal case is a challenging task.
The shared circuits model (SCM):
How control, mirroring, and
simulation can enable imitation,
deliberation, and mindreading
Susan Hurley
Department of Philosophy, University of Bristol, Bristol BS8 1TB, and All Souls
College, Oxford University, Oxford OX1 4AL, United Kingdom
Susan.hurley@bristol.ac.uk
http://eis.bris.ac.uk/~plslh/
Abstract: Imitation, deliberation, and mindreading are characteristically human sociocognitive skills. Research on imitation and its role
in social cognition is flourishing across various disciplines. Imitation is surveyed in this target article under headings of behavior,
subpersonal mechanisms, and functions of imitation. A model is then advanced within which many of the developments surveyed
can be located and explained. The shared circuits model (SCM) explains how imitation, deliberation, and mindreading can be
enabled by subpersonal mechanisms of control, mirroring, and simulation. It is cast at a middle, functional level of description, that
is, between the level of neural implementation and the level of conscious perceptions and intentional actions. The SCM connects
shared informational dynamics for perception and action with shared informational dynamics for self and other, while also showing
how the action/perception, self/other, and actual/possible distinctions can be overlaid on these shared informational dynamics. It
avoids the common conception of perception and action as separate and peripheral to central cognition. Rather, it contributes to
the situated cognition movement by showing how mechanisms for perceiving action can be built on those for active perception.
The SCM is developed heuristically, in five layers that can be combined in various ways to frame specific ontogenetic or phylogenetic
hypotheses. The starting point is dynamic online motor control, whereby an organism is closely attuned to its embedding environment
through sensorimotor feedback. Onto this are layered functions of prediction and simulation of feedback, mirroring, simulation of
mirroring, monitored inhibition of motor output, and monitored simulation of input. Finally, monitored simulation of input
specifying possible actions plus inhibited mirroring of such possible actions can generate information about the possible as opposed
to actual instrumental actions of others, and the possible causes and effects of such possible actions, thereby enabling strategic
social deliberation. Multiple instances of such shared circuits structures could be linked into a network permitting decomposition
and recombination of elements, enabling flexible control, imitative learning, understanding of other agents, and instrumental and
strategic deliberation. While more advanced forms of social cognition, which require tracking multiple others and their multiple
possible actions, may depend on interpretative theorizing or language, the SCM shows how layered mechanisms of control,
mirroring, and simulation can enable distinctively human cognitive capacities for imitation, deliberation, and mindreading.
Keywords: action; active perception; control; embodied cognition; imitation; instrumental deliberation; isomorphism; mindreading;
mirroring; mirror neurons; shared circuits; simulation; social cognition
BEHAVIORAL AND BRAIN SCIENCES (2008) 31, 1–58
Printed in the United States of America
doi: 10.1017/S0140525X07003123
# 2008 Cambridge University Press 0140-525X/08 $40.00 1
2. 1. Introduction: From active perception to social
cognition and beyond
Like many today, I view perception as inherently active
(Hurley 1998; Hurley & Noë 2003, O’Regan & Noë
2001a; 2001b; Noë 2004) and cognition as embodied and
situated. How does cognition relate to active perception?
This article shows how subpersonal resources for social cog-
nition can be built on those for active perception. Its central
issue is the following: How is it possible to perceive events
as instrumentally structured intentional actions, and to
learn new instrumental actions by means of such percep-
tions of actions? The article shows how the structures and
mechanisms for perceiving action and for situated social
cognition can be built on those for active perception. It
extends my previous contributions to a view of perception
as inherently active and of cognition as embodied and situ-
ated (Hurley 1998; Hurley & Noë 2003).
The classical sandwich conception of the mind –
widespread across philosophy and empirical sciences of
the mind – regards perception as input from world to
mind, action as output from mind to world, and cognition
as sandwiched between. I have argued that the mind isn’t
necessarily structured in this vertically modular way
(Brooks 1999; Hurley 1998). Moreover, there is growing
evidence that it is not actually so structured in specific
domains, where perception and action share dynamic
information-processing resources as embodied agents
interact with their environments, rather than functioning
as separate buffers around domain-general central
cognition.
Rather, cognitive resources and structure can emerge,
layer by layer, from informational dynamics, enabling
both perception and action. Such a horizontally modular
structure can do significant parts (I don’t claim all) of
the work the classical sandwich conception assigned to
central cognition. Here I show how this promise can be
fulfilled for the perception of action and associated social
cognition, as embodied agents interact with their social
environments.
I first review recent work on social cognition, focusing
on imitation (Hurley 2005b; Hurley & Chater 2005a;
2005b). Imitation is still popularly regarded as cognitively
undemanding. However, Thorndike (1898) showed that
many animals can learn through individual trial and
error but not imitatively; scientists regard the later as
more cognitively demanding. Imitative ability is rare
across animal species and linked to characteristically
human capacities: for language, culture, and understanding
other minds (Arbib et al. (2000); Arbib 2005; Arbib &
Rizzolatti 1997; Barkley 2001; Frith & Wolpert 2004;
Gallese 2000; 2001; 2005; Gallese & Goldman 1998; Gallese
et al. 2004; Gordon 1995b; Iacoboni 2005; Meltzoff 2005;
Rizzolatti & Arbib 1998; 1999; Stamenov & Gallese 2002;
Tomasello 1999; Whiten et al. 2005b; Williams et al. 2001).
Imitation is important in adult human sociality, as well as
human development, in ways we’re just beginning to
understand.
Part 1 of this article reviews recent research on imita-
tion, under the headings of behavior, subpersonal mech-
anisms, and functions. Part 2 presents a functional
architecture that shows how subpersonal mechanisms of
control, mirroring, and simulation can enable distinctively
human skills of imitation, deliberation, and action under-
standing. The shared circuits model (SCM) draws together
many threads of work from Part 1. It includes elements
suggested by various researchers, contributes further
elements, and unifies these in a distinctive framework.
SCM aims to show how the following are possible:
SUSAN HURLEY, who passed away in August 2007, had
been Professor and Chair in Philosophy at the University
of Bristol since August 2006 and was also a Fellow of All
Souls College in Oxford. She had served as Professor and
Politics and International Studies affiliate at the Univer-
sity of Warwick for the previous twelve years. Her most
recent research had been in philosophy of psychology
and neuroscience, focusing on consciousness, social cog-
nition (imitation and mindreading), and action (ration-
ality, control, responsibility). She had also worked in
political philosophy and related areas, with a particular
interest in bringing the cognitive and social sciences
into constructive contact. Hurley’s books include
Natural Reasons: Personality and Polity (Oxford Univer-
sity Press, 1989), Consciousness in Action (Harvard Uni-
versity Press, 1998), Justice, Luck and Knowledge
(Harvard University Press, 2003), and edited volumes
on the foundations of decision theory, on imitation, on
rationality in animals, and on human rights. Hurley did
her undergraduate work in philosophy at Princeton Uni-
versity and her graduate work in philosophy (a B.Phil.
and a doctorate) at Oxford University. She also earned
a law degree at Harvard University. After four years
(1981–1984) as a Junior Research Fellow at All Souls
College, Oxford, Hurley spent ten years as a Tutorial
Fellow in Philosophy at Oxford, before moving to
Warwick. At the time of her passing, Hurley was one of
the Principal Investigators on a large multicentre
project studying the role of the natural and social
environment in shaping consciousness.
ANDY CLARK is Professor of Philosophy in the School
of Philosophy, Psychology, and Language Sciences, at
Edinburgh University in Scotland. He was a close
friend to Susan Hurley, whose ideas concerning mind
and dynamics have had a large influence on his own
work which concerns the nature of mind, and the cog-
nitive role of bodily and environmental structures and
processes. He is the author of several books including
Being There: Putting Brain, Body And World Together
Again (MIT Press, 1997), Natural-Born Cyborgs:
Minds, Technologies And The Future Of Human Intel-
ligence (Oxford University Press, 2003) and Supersiz-
ing the Mind: Embodiment, Action and Cognitive
Extension (forthcoming with Oxford University Press).
JULIAN KIVERSTEIN is a Postdoctoral Research Fellow
in the Department of Philosophy at the University
Edinburgh working as part of the collaborative
research project CONTACT – Consciousness in Inter-
action. CONTACT is a part of the EUROCORES
program, Consciousness in the Natural and Cultural
Context, which will run from 2006 to 2009. Susan
Hurley was a Principle Investigator of the CONTACT
group based at Bristol University. Kiverstein and
Hurley were just beginning to collaborate on a
project investigating the relationship between social
cognition and consciousness using the shared circuits
model, one of the many applications of shared circuits
Hurley had planned to explore. Kiverstein plans to
complete this work as part of postdoctoral research.
Hurley: The shared circuits model
2 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
3. 1. Building subpersonal informational resources for
situated social cognition on those for active perception,
while
2. Uniting a large body of evidence and theorizing in a
common framework;
3. Avoiding the “classical sandwich”; and
4. Respecting the personal/subpersonal distinction and
avoiding interlevel isomorphism assumptions.
Philosophers distinguish descriptions of contentful
actions and mental states of persons from subpersonal
(informational or neural) descriptions (Bermudez 2000;
2003; Dennett 1969; 1991; Elton 2000; Hornsby 2000;
McDowell 1994). At the subpersonal level of description,
information is processed and the cycling of causes and
effects knits actively embodied nervous systems into
environments they interact with.1
But these processes
are not correctly attributed to persons. Persons see trees,
make friends, look through microscopes, vote, want to
be millionaires. Subpersonal informational and causal the-
ories explain how personal-level phenomena become pos-
sible – are enabled – but need not share structure with
personal-level descriptions of processes as rational or con-
scious.2
I distinguish three levels of description: the per-
sonal level, the informational/functional subpersonal
level, and the neural subpersonal level.
Two questions arise about personal/subpersonal
relations: (1) How are specific personal-level capacities
actually enabled by subpersonal processes? (2) What
kinds of subpersonal processes could possibly do the
enabling work? For example, must there be isomorphism
between levels? Views about question 2 can influence
answers to question 1.
SCM addresses question 2 for social cognition by using
subpersonal resources from an active perception
approach. SCM is cast at the subpersonal functional
level, not the personal or the neural levels, though it
aims both to show how certain personal-level capacities
can be informationally enabled and to raise empirical
questions about neural implementation. Since SCM
addresses the “how possibly?” rather than the “how actu-
ally?” question, it provides a higher-order theoretical
model. But it also provides generic heuristic resources
for framing specific first-order hypotheses and predictions
about specific ontogenetic or phylogenetic stages. Its five
layers, detailed below, can be re-ordered in formulating
specific first-order hypotheses.
SCM’s central hypothesis is that associations underwrit-
ing predictive simulation of effects of an agent’s own
movement, for instrumental control functions, can also
yield mirroring and “reverse” simulation of similar per-
ceived movements by others. Mirroring allows ends/
means associations with instrumental control functions to
be accessed for simulative functions bilaterally, so causes
of observed movements can be simulated, as well as
effects of intended acts. Such bilaterally accessible simu-
lations of instrumental structure can provide enabling
information for deliberation, imitation, and understanding
the instrumental acts of others. Shared dynamics for action
and perception can provide the foundations of shared
dynamics for self and other, and of the self/other and
actual/possible distinctions characteristic of human
cognition.
Simulation has a generic sense throughout, including,
but broader than, that in simulationist theories of
mindreading (Gallese 2003; see Goldman 1989; 1992 on
process-driven simulation). Simulation uses certain pro-
cesses to generate related information, rather than theoriz-
ing about them in separate meta-processes. Effects or
causes can be simulated, online or offline. Simulation
can be subpersonal or personal; in SCM it is subpersonal.
Subpersonal processes that predict results of movement
online can also generate information about results of poss-
ible movements offline. Subpersonal mirroring that
enables copying can also generate information covertly
about observed movements or their goals, without overt
copying (cf. Barkley [2001] on executive functions as
covert behavior).
2. Part 1. Review
I begin in Part 1 by reviewing recent work, strands of
which are knitted together by SCM in Part 2.
2.1. Behavior
Imitative learning is a sophisticated form of social cogni-
tion. It requires copying in a generic sense: Perception
of behavior causes similar behavior by an observer, and
the similarity plays a role – not necessarily conscious-
ly – in generating the observer’s behavior. True imitation,
restrictively understood, requires novel action learned by
observing another do it, plus instrumental or means/
ends structure: the other’s means of achieving her goal is
copied, not just her goal or just her movements. The
concept of true imitation is contested, given the different
aims and methodologies of imitation researchers (Byrne
2005; Heyes 1996; 2001; Rizzolatti 2005).
Other forms of social learning can seem similar to imita-
tion, but should be distinguished. In stimulus enhance-
ment, another’s action draws your attention to a
stimulus, which triggers an innate or previously learned
response; but a novel action isn’t learned directly from
observation. Bird A’s pecking may draw bird B’s attention
to a food, which evokes pecking in bird B. In goal emula-
tion,3
you observe another achieving a goal by certain
means, find that goal attractive, and try to achieve it your-
self. Monkey A may use a tool in a certain way to obtain an
attractive object, leading monkey B to acquire the goal of
obtaining a similar object. Through his own trials and
errors, monkey B may arrive at the same type of tool use
to obtain the object. Emulation is found in macaques,
who have not shown imitative learning. In movement
priming, bodily movements are copied, but not as
learned means to a goal. Primed movements can be
innate, as in contagious yawning.
Goal emulation and movement priming provide the
ends-and-means components of full-fledged imitation.
Ends and means can be relatively distal or proximal; the
distinction is relative, not absolute. Misunderstandings
can result concerning whether ends, means, or both are
copied and hence whether imitation or emulation is
present (Voelkl & Huber 2000, pp. 196, 201). A movement
can be the proximal means to a bodily posture, which is
regarded as the proximal end of the movement (Graziano
et al. 2002, pp. 354–55), but posture can also be a means
to more distal ends – effects on objects or others in social
groups. Complex imitation can involve structured
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 3
4. sequences or hierarchies of ends–means relations: acquir-
ing a goal, learning to achieve it via subgoals, and so on.
How are these forms of copying distributed across
animals, children, and adults? Stimulus enhancement,
goal emulation, and movement priming are certainly
found in nonhuman animals. But careful experiments are
needed to distinguish these from imitation proper and
obtain evidence of the latter. The two-action paradigm
has been the tool of choice. Suppose two demonstrators
obtain an attractive result by two different means:
One group of animals observes one demonstrator, and
the other observes the other demonstrator. Will the
observer animals tend differentially to copy the specific
method they have seen demonstrated? If not – if the
animals’ choices of method do not reflect the specific
method they have observed, say, because both groups con-
verge on one method – they may be displaying stimulus
enhancement or goal emulation plus trial-and-error learn-
ing, but not imitative learning. Even if they do
differentially display the behavior demonstrated, this
may be merely movement priming if the behavior is
already in their repertoire. But if the behavior is differen-
tially used in a new way to achieve a result, it expresses imi-
tative learning (Call & Tomasello 1994; Nagell et al. 1993;
Voelkl & Huber 2000).
The difference between copying ends and copying
means is important for theorizing the phylogeny of imita-
tion and action understanding. (Action understanding is
short for understanding observed behavior as goal-
directed action.) Is action understanding phylogenetically
prior to imitation? This view seems to face an objection:
Some animals copy movements (schooling fish), though
we don’t think they understand the others’ actions.
A response to this objection distinguishes movement
copying from mirroring of goals (Rizzolatti 2005) and
both from imitation. Movement copying may precede
action understanding, whereas action understanding may
require goal mirroring, but precede imitation. True imita-
tion involves something phylogenetically rare: the flexible
interplay of copying ends and copying means; a given
movement can be used for different ends and a given
end pursued by various means (Barkley 2001, p. 8;
Tomasello 1999). This is something humans are distinc-
tively good at.
It is difficult to find evidence of true imitation in nonhu-
man animals (Byrne 1995; Galef 1988; 1998; 2005; Heyes &
Galef 1996; Tomasello 1996; Tomasello & Call 1997;
Voelkl & Huber 2000; Zentall 2001). Early work with
chimps seems to reveal imitation, but critics have chal-
lenged this interpretation effectively; the results of sub-
sequent experiments were negative for chimp imitation.
Skeptics about nonhuman imitation long had the upper
hand; for example, Tomasello et al. (1993) found no con-
vincing evidence of nonhuman imitative learning. They
proposed that understanding behavior as intentional dis-
tinguishes human from nonhuman social learning. On
this view, humans can imitate observed means or choose
other means to emulate observed goals. Other animals
don’t distinguish means and goals this way; rather, they
copy movements without understanding their relevance
to goals, or learn about the affordances of objects by
observing action on them. In neither case, it was
claimed, do other animals learn about the intentional,
means/end structure of observed action.
Many skeptics have now been won over by work on imi-
tation in great apes and monkeys (Voelkl & Huber 2000;
Whiten et al. 2005b), dolphins (Herman 2002), and birds
(Akins & Zentall 1996; 1998; Akins et al. 2002; Hunt &
Gray 2003; Pepperberg 1999; 2002; 2005; Weir et al.
2002). Continuities are described along a spectrum from
the capacities of other social animals to human socio-
cognitive capacities (Arbib 2005; Tomasello 1999). For
example, innovative experiments extend the two-action
method by using “artificial fruits” that can be opened in
different ways to obtain a treat: Chimps tend to imitate
for one aspect of a demonstrated task and emulate for
another aspect, whereas children tend to imitate both
aspects, even when the method imitated is inefficient.
These and other experiments suggest that chimps imitate
more selectively than children (Whiten 2002; Whiten
et al. 1996; 2005b; see also Call & Tomasello 1994; Galef
2005; Harris & Want 2005; Heyes 1998; Nagell et al.
1993; Tomasello & Carpenter 2005).
Children have been called “imitation machines”
(Tomasello 1999, p. 159). They don’t always imitate unse-
lectively and sometimes emulate goals instead (Gergely
et al. 2002), but they have a greater tendency than
chimps to imitate rather than emulate when the method
demonstrated is transparently inefficient or futile
(Tomasello 1999, pp. 29–30). After seeing a demonstrator
use a rake inefficiently, prongs down, to pull in a treat,
2-year-old children do the same; they almost never turn
the rake over to use more efficiently, edge down. By con-
trast, chimps given a parallel demonstration, tend to try
turning the rake over (Nagell et al. 1993).4
The differential
tendency of children and chimps to imitate suggests an
interplay of biological and cultural influences, with a role
for innate endowment enabling human imitation
(perhaps a matter of articulated relations among multiple
mirror subsystems, enabling recombinant structure in
social learning, rather than the presence versus absence
of a mirror system at all; see discussion of mechanisms
in sect. 2.2).
Imitative and related behaviors appear throughout
human development (Meltzoff 1988a; 1990; 1995; 1996;
2002a; 2002b; 2005; Meltzoff & Moore 1977; 1983a;
1983b; 1989; 1997; 1999; 2000). Infants younger than
1 month of age appear to copy facial gestures. By 14
months, infants imitate a novel act a week later: They
turn on a light by touching a touch-sensitive panel with
their forehead instead of their hand, differentially
copying the novel means demonstrated, as well as the
result (Meltzoff 1988a; 2005; cf. Gergely et al. 2002).
They don’t turn the light on in this odd way unless they
have seen it demonstrated. By 15 to 18 months, infants
recognize the underlying goal of an unsuccessful act they
observe, and produce it: After seeing an adult try but fail
to pull a dumbbell apart in her hands, they succeed in
pulling it apart using knees and hands. But they don’t
pick up goals from failed “attempts” involving similar
movements by inanimate devices, thus apparently discri-
minating agents from non-agents (Meltzoff 1988a; 1995;
1996; 2005; Meltzoff & Moore 1977; 1999; Tomasello &
Carpenter 2005). Children’s perception of behavior
tends to be enacted automatically in similar behavior,
unless actively inhibited; but frontal inhibitory functions
are not well developed in young children (Barkley 2001,
pp. 5, 22; Kinsbourne 2005; Preston & de Waal 2002, p. 5).
Hurley: The shared circuits model
4 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
5. Adult “imitation syndrome” patients with frontal brain
lesions also imitate uninhibitedly (Barkley 2001, p. 15;
Frith 1992, pp. 85–86; Lhermitte 1983; 1986; Lhermitte
et al. 1986, p. 330). They persistently copy the exper-
imenter’s gestures, though not asked to, even when
these are socially unacceptable or odd, such as putting
on eyeglasses when already wearing glasses.
But the human copying tendency isn’t confined to the
young or brain damaged! Normal adults can usually
inhibit overt imitation selectively, which is evidently adap-
tive, but their underlying tendency to copy is readily
revealed or released. Overt imitation is the disinhibited
tip of the iceberg of continual covert imitation (Barkley
2001; Dijksterhuis 2005). Experiments show how action
is modulated or induced by perception of similar action
(Brass et al. 2001; Prinz 2002). Imitative tasks have
shorter reaction times than nonimitative tasks; gestures
are faster when participants are primed by perceiving
similar gestures or their results or goals – even when
primes are logically irrelevant to the task (W. Prinz
2005). Similarity between stimulus and response also
affects which response is made. Normal adults, instructed
to point to their nose when they hear “Nose!” and to a
lamp when they hear “Lamp,” performed perfectly while
watching the experimenter demonstrate the required per-
formance, but made mistakes when watching the exper-
imenter doing something else: they tended to copy what
they saw done rather than to follow instructions
(Eidelberg 1929; Prinz 1990). Movements can be induced
by actions you actually perceive or by actions you would
like to perceive – as when moviegoers or sports fans in
their seats make movements they would like to see
(W. Prinz 2005). Visually or verbally represented, as well
as observed, actions can induce similar actions.
It is helpful to distinguish copying of specific behaviors
from chameleon effects, where complex patterns of beha-
vior are induced – a relevant kind of copying, if not
strict imitation. In an experiment involving specific beha-
viors, when normal adults interact in an unrelated task
with someone rubbing her foot, they rub their own feet
significantly more. Transferred to another partner who
touches his face, they touch their own faces instead. Dem-
onstrations of chameleon effects show that exposure to
traits and stereotypes automatically elicits general patterns
of behavior and attitude and influences how behavior is
performed (Bargh 1999; 2005; Bargh & Chartrand 1999;
Bargh et al. 1996; 2001; Chartrand & Bargh 1999;
Chartrand et al. 2005; Dijksterhuis & Bargh 2001).
Normal adults primed with stimuli associated with traits
(e.g., hostility, rudeness, politeness) or stereotypes (e.g.,
elderly persons, college professors, soccer hooligans)
tend to behave in accordance with the primed traits or
stereotypes. For example, hostility-primed participants
deliver more intense “shocks” than control participants
in subsequent, ostensibly unrelated experiments based
on Milgram’s (1963) classic shock experiments. Priming
can also affect intellectual performance: College pro-
fessor–primed participants perform better and soccer
hooligan–primed participants perform worse than con-
trols on a subsequent, ostensibly unrelated general
knowledge test (Dijksterhuis 2005; Dijksterhuis & van
Knippenberg 1998).
Such priming results are robust across a wide range of
verbal and visual primes and induced behavior, using
dozens of stereotypes and general traits, and various
priming methods, including primes perceived consciously
and subliminally. Whether subjects perceive primes
consciously or not, they are unaware of any influence or
correlation between primes and their behavior. These
influences are rapid, automatic, and unconscious, apply
both to goals and means, and don’t depend on subjects’
volition or having independent goals that would rationalize
their primed behavior.
Copying, at various levels of generality, is thus a default
social behavior for normal human adults; it requires
specific overriding or inhibition (Barkley 2001, p. 22;
Dijksterhuis 2005; Preston & de Waal 2002). Just thinking
about or perceiving action automatically increases, in ways
participants are unaware of, the likelihood that they will
perform similar actions themselves. Nevertheless, these
influences are often inhibited, as when goals make con-
flicting demands: elderly–primed participants tend to
walk more slowly, but not if they have independent
reasons to hurry.
2.2. Mechanisms
Copying perceived behavior seems to pose a correspon-
dence problem (Nehaniv & Dautenhahn 2002): How is
another’s observed action translated into the observer’s
similar performance? When I copy your hand movements,
I can see my own hands, though my visual perspectives on
the two movements are different. But when I copy your
facial expressions, I cannot see my own face. What infor-
mation and mechanisms are needed to map perception
to similar behavior?
Evidence that newborns copy facial gestures, though
they cannot see their own faces, suggests innate supramo-
dal correspondences between action and perception of
similar action (Meltzoff 1988a; 1990; 1995; 1996; 2002a;
2002b; 2005; Meltzoff & Moore 1977; 1983a; 1983b;
1989; 1997; 1999; 2000). Although further correspon-
dences could be acquired as imitative abilities develop,
skeptics about newborn copying can also be skeptical
about the need to postulate any innate correspondences
(Anisfeld 1979; 1984; 1991; 1996; 2005; Anisfeld et al.
2001; Heyes 2005).
Heyes, who is one such skeptic, argues that sensorimo-
tor associations subserving copying can be acquired
through general-purpose associative learning mechanisms
whereby neurons that fire together wire together. Direct
sensorimotor associations between motor output and
sensory feedback could result from watching one’s own
hand gestures. An indirect route is needed when the
agent cannot perceive her own actions, as in facial
expressions: The sensorimotor association could be
mediated by environmental items such as mirrors, action
words, or stimuli that evoke similar behavior in the actor
and in other agents the actor observes. Moreover, adults
commonly copy infants, performing the associative func-
tion of a mirror. When baby smiles and father smiles
back, baby’s motor output is associated with sensory
input from father’s smile (Heyes 2005, p. 161; Preston &
de Waal 2002, p. 8). Imitation can thus develop from inter-
actions between organisms with associative learning mech-
anisms and certain cultural environments (see Heyes
2001; 2005; see also SCM, Layer 3 in sect. 3.3 here).
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 5
6. Common coding for perception and action has been
postulated to explain human copying tendencies. On this
view, perception and action share subpersonal processes
carrying information about (“coding for”) what is perceived
or intended, in which perception and action are not distin-
guished. The differentiation between perception and
action is overlaid on those shared resources, so that they
are informationally interdependent at a basic level. If
capacities X and Y share an information space, their com-
monality is informationally prior to their differentiation.
Meltzoff and Moore (1997) postulate common coding of
perception and action in explaining infant imitation: pro-
prioceptive feedback is compared to an observed target
act, where these are coded in common, supramodal
terms. Innate common coding could initially be for
relations among, say, lips and tongue; more dynamic,
complex, and abstract common coding could develop
with experience of body babbling. Common coding
might also be acquired as in Heyes’s (2005) model.
Wolfgang Prinz (1984; 2005; cf. Bekkering &
Wohlschläger (2002); Preston & de Waal 2002, pp. 4ff.,
9–10) appeals to common coding of perception and
action to explain the normal adult tendency to imitate
and the reaction-time advantage of imitative tasks.
Common coding facilitates imitation, avoiding the corre-
spondence problem and any need for translation
between unrelated input and output codes to solve it.
Prinz associates common coding with what William
James called ideomotor theory, on which every represen-
tation of movement awakes in some degree the movement
it represents (Brass 1999; Prinz 1987). Perceiving
another’s observed movement tends inherently to
produce similar movement by the observer, and primes
similar movement even when it doesn’t break through
overtly. Regular concurrence of action with perceived
effects allows prediction of an action’s effects and selection
of action, given an intention to produce certain effects
(Greenwald 1970; 1972). Thus, representation of an
action’s regular result, whether proximal or distal, can
evoke similar action, in the absence of inhibition.
Other sources also support the view that perception and
action share processing resources. Observing an action
primes the very muscles needed to perform the same
action (Craighero et al. 2002; Fadiga et al. 1995; 2002).
Watching an action sequence speeds the observer’s per-
formance of that sequence; merely imagining a skilled
performance, in sport or music, improves performance – is
a way of practicing – as many athletes and musicians know
(Jeannerod 1997, pp. 117, 119–22; Pascual-Leone 2001).
Similar points concern perception and experience of
emotion: Gordon argues that a special containing mechan-
ism, which isn’t fail-safe, is needed to keep emotion recog-
nition from producing emotional contagion. On his
simulationist theory, only a thin line separates one’s own
mental life from one’s representation of another’s; offline
representations of others tend inherently to go online
(Gordon 1995b; cf. Adolphs 2002; Preston & de Waal 2002).
Common coding theories characterize subpersonal
architectures for copying functionally. What neural pro-
cesses might implement such functional architectures?
Certain neurons directly link perception and action:
their firing correlates with specific perceptions and
specific actions. Canonical neurons (Gallese 2005;
Rizzolatti 2005) reflect affordances (Iacoboni 2005; Miall
2003): They fire when an animal perceives an object that
affords a certain type of action and when the animal per-
forms the afforded action. Mirror neurons fire when an
animal perceives another agent performing a type of
action, and also when the animal performs that type of
action itself; they don’t distinguish own action from
others’ similar actions (see SCM, Layer 3 in sect. 3.3).
Some fire, for example, when a monkey sees the exper-
imenter bring food to her own mouth with her hand or
when the monkey brings food to his own mouth with his
hand (even in the dark, so the monkey cannot see his
hand). Specificity of tuning varies.
How mirror neurons relate to imitation is of much
current interest (e.g., see Frith & Wolpert 2004; Rizzolatti
et al. 2002; Williams et al. 2001). It may be tempting to
think they avoid correspondence problems, thus facilitat-
ing imitation: If the same neurons code for perceived
action and similar performance, no translation is needed.
But things are not so simple. Rizzolatti, one of the disco-
verers of mirror neurons, holds that imitation requires
both the ability to understand another’s action and the
ability to replicate it. On his view, recall, action under-
standing precedes imitation phylogenetically; action
understanding is subserved by mirror systems, which
might be necessary, but are not sufficient, for imitation.
Rizzolatti (2005) suggests that the motor resonance set
up by mirror neurons makes action observation meaning-
ful by linking it to the observer’s own potential actions.
Mirror neurons were discovered by single-cell record-
ing in macaques (Di Pellegrino et al. 1992; Rizzolatti
et al. 1988; 1995), which can emulate but have not been
shown to imitate in a strict sense (cf. Voelkl & Huber
[2000] on marmoset imitation). Evidence for human
mirror systems (Craighero et al. 2002; Decety &
Chaminade 2005; Decety et al. 1997; Fadiga et al. 2002;
Hari et al. 1998; Iacoboni et al. 2005; Rizzolatti et al.
1996; Ruby & Decety 2001) includes brain imaging and
demonstrations that observing another person move
primes the muscles needed to move in a similar manner
(whether or not movements are goal directed; Fadiga
et al. 1995).
Rizzolatti (2005) describes mirror neurons in monkey
frontal brain area F5 as part of a circuit including parietal
area PF and visual area STS (superior temporal sulcus).
He regards a similar human brain circuit as a control
system: Sensory results associated with certain movements
are compared in PF to observed target movements,
enabling imitative learning (cf. Iacoboni 2005, with
regard to locating the comparator in STS). Differently
structured mirror systems may explain different copying
capacities across species. In monkeys, mirror neurons
appear to code for the goals or results of performed or
observed actions.5
By contrast, human mirror systems
include specific movements that can be means to achiev-
ing goals (Fadiga et al. 1995). Recall how the difference
between mirroring ends versus means of action matters
for the view that action understanding precedes imitation
phylogenetically. If seeing someone reach for an apple
produces motor activation associated with the same goal
in the observer (though not necessarily with the same
movements in the observer), that could provide infor-
mation about the observed action’s goal directness. But
it wouldn’t provide information about how to achieve the
goal by means of the observed movements, as in imitation.
Hurley: The shared circuits model
6 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
7. Brain imaging suggests a division of labor in the human
mirror system: Its frontal regions tend to code for goals of
action, whereas its parietal regions tend to code for means
(i.e., movements; Iacoboni 2005). (Monkey parietal mirror
neurons seem to be goal–related; Fogassi et al. 2005;
Nakahara & Miyashita 2005.) One theory of how this div-
ision of labor enables imitation relates signals generated by
these brain regions to comparator circuits for instrumental
motor control combining inverse and forward models.
Inverse models estimate the motor plan needed to
achieve a goal in a given context. They can be adjusted
by comparison with real motor feedback, but this is slow.
It is often more efficient to use real feedback to train
forward models, which anticipate the sensory effects of
motor plans, associating action with its perceived results
(as do mirror neurons). Forward models combine with
inverse models to control goal-directed behavior more effi-
ciently. Forward models can predict the results of imita-
tive motor plans for comparison to observed action, and
motor plans can be adjusted until a match obtains
(Flanagan et al. 2003; Iacoboni 2005; Miall 2003;
Wolpert et al. 2003).
Thus, mirror neurons are arguably part of the neural
basis for true imitation, though not sufficient for it.
Monkey mirror neurons code for ends rather than
means. Human mirror systems, by contrast, have articu-
lated structure: some regions code for goals, whereas
others code for specific movements that are means to
goals. It has been suggested that human mirror systems
enable imitation (not just emulation) because they code
for means as well as ends (unlike the macaque’s system),
and that mirror neurons contribute predictive forward
models to subpersonal comparator control circuits.
2.3. Functions
Human brains differ most from chimp brains in expanded
areas around the Sylvian fissure which subserve imitation,
language, and action understanding – where many mirror
neurons are found (Iacoboni 2005). Can mirror systems
illuminate the functions of imitation in relation to distinc-
tively human capacities – for language, or for identifying
with others and understanding the mental states motivat-
ing their actions? The relationships among capacities for
imitation, language, and mindreading are important for
understanding phylogeny and human development. Does
development of either language or mindreading depend
on imitation? If so, at what levels of description and in
what senses of “depend”? Or does dependence run the
other way? Or both ways, dynamically? Answers may
differ for language and for mindreading. Issues about
relations between imitation and mindreading entwine
with issues about whether mindreading is best understood
as theorizing about other minds or as simulating them.
I shall survey some hypothesized functions of imitation
in language, cultural evolution, cooperation, and mind-
reading. The first three topics, discussed briefly, provide
context for SCM and illustrate its broader relevance to
understanding what is distinctive about human minds.
Mindreading is directly related to SCM and so is discussed
more fully.
2.3.1. Language. It has been suggested that “mirror
neurons could .. . be an important neural stepping stone .. .
to spoken language” (Miall 2003, p. 1). Mirror systems
for action goals include Broca’s area,6
a main language
area of human brains, which is active during imitative
tasks. Moreover, transient virtual “lesions” to Broca’s
created by transcranial magnetic stimulation interfere
with imitative tasks (Heiser et al. 2003; Iacoboni 2005).
Nativism about language might view Broca’s as the best
candidate for an innate language module (M. Iacoboni,
in discussion). But discovery that Broca’s subserves
mirror systems and has some role in enabling imitation
has generated new arguments about how language acqui-
sition could build on capacities for action understanding
and imitation, in either evolutionary or developmental
time frames, exploiting imitative learning rather than, or
in addition to, innate linguistic knowledge (Arbib 2005;
Arbib & Rizzolatti 1997; Iacoboni 2005; Rizzolatti &
Arbib 1998; 1999; Stamenov & Gallese 2002). (On
language and social learning, see Baldwin 1995; Barkley
2001; Christiansen 1994; 2005; Christiansen & Kirby
2003; Deacon 1997; on establishing shared reference to
objects through joint attention, via gaze following and
role-reversal imitation, see Tomasello 1999.)
What features of imitation and human mirror systems
might language build on? First, I suggest, flexible articu-
lated relations between means and ends in imitative learn-
ing could be an evolutionary precursor of arbitrary
relations between symbols and referents. Decoupling a
particular bodily movement from a given result and treat-
ing it as a potential means to various possible results in
varying circumstances (see SCM, Layers 2 plus 4, in
sects. 3.2 and 3.4) may be a step toward treating it as
lacking an intrinsic function and so available for an arbitra-
rily or conventionally assigned communicative function.
Second, mirror systems provide a common code for
actions of self and other, and thus for language production
and perception; by enabling intersubjective action under-
standing, mirror systems may be the basis for the intersub-
jective parity, or sharing of meaning, essential to language
(Arbib 2005; Iacoboni 2005).
Third, the flexible recombinant structure of ends and
means in imitation may be a precursor of recombinant
grammatical structure in language (Arbib 2005). The
latter may result when creatures with recombinant imita-
tive skills learn to pursue their goals by recombinant
manipulation of external symbols.
Fourth, finding recombinant units of action in streams
of bodily movement has parallels with finding linguistic
units (e.g., words) in continuous acoustic streams of
speech (Byrne 2005). The modular structure of skilled
action facilitates flexible recombination. Patterns of
action organization could be learned in program-level
imitation,7
despite variation in implementational details,
by using mirror mechanisms plus mechanisms for
parsing behavior modules. Behavior parsing and the
recombinant structure of program-level imitation may be
precursors of human capacities to perceive underlying
structures of intentions or causes in the surface flux of
experience – and perhaps of syntactic parsing and the
recombinant structure of language.
2.3.2. Cultural evolution. A more fundamental question
is: Why might evolution favor neural structures that
enable various forms of copying to begin with? Suppose
individuals vary in behavioral traits that are not genetically
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 7
8. heritable, so some reproduce more successfully than
others. Their offspring may benefit by acquiring behaviors
from their successful parents by copying, as well as geneti-
cally. By copying reproductively successful parents, off-
spring can acquire nonheritable behaviors associated
with appropriate environmental conditions. If individual
learning is costly, copying may contribute more to
genetic fitness.
If true imitation requires mirror circuits for means and
ends to be linked in ways that give social learning recom-
binant flexibility, it should be harder to evolve than move-
ment priming or emulation. And, indeed, it is found in
fewer species. But wouldn’t this rare development from
other forms of copying to imitation be maladaptive?
Recall the short-term disadvantage of children compared
to chimps in two-action paradigms: Children have a
greater tendency to imitate even inefficient models,
whereas chimps have a greater tendency to emulate and
find more efficient means to attractive goals (Nagell
et al. 2003; Whiten et al. 2005b). Despite this, could the
stronger imitative tendency be adaptive long-term?
Yes: via the ratchet effect (Tomasello 1999). Gifted or
lucky individuals may discover efficient new means to
goals – means that are not readily rediscoverable by inde-
pendent trial-and-error learning. These would be lost
without recombinant imitative learning, which preserves
and disseminates valuable instrumental innovations, pro-
viding a platform for further innovation. Once imitation
evolves genetically, it provides a mechanism of cultural
and technological transmission, accumulation, and evol-
ution. The effects of imitative copying and selection inter-
twine with those of genetic copying and selection; culture
and life coevolve (see and cf. Baldwin 1896; Barkley 2001,
p. 21; Blackmore 1999; 2000; 2001; Boyd & Richerson
1982; 1985; Dawkins 1976/1989; Deacon 1997; Dennett
1995; Gil-White 2005; Henrich & Boyd 1998; Henrich &
Gil-White 2001; Hurley & Chater 2005a, part 4).
The capacity for selective imitation may have an import-
ant role in underwriting the ratchet effect (Harris & Want
2005). Imitation with selective inhibition has the advan-
tages of theft over honest toil: Instead of letting hypotheses
die in his stead, a selective imitator lets others die in his
stead, reaping the benefits of success without unusual
native wit while avoiding the costs of trial and error. Imi-
tative social environments may in turn generate pressure
to prevent successful techniques being appropriated cost
free by competitors, resulting in capacities for covert or
simulated action, shielded from potential imitative theft
(Barkley 2001, pp. 9, 18–21).
2.3.3. Cooperation. As well as being subject to automatic
copying influences, humans often deliberately select
a behavior pattern to imitate because it is associated
with certain traits or stereotypes, even if they themselves
don’t exemplify these traits or stereotypes. This can be
benign and contribute to moral development (J. Prinz
2005); perhaps I can become virtuous, as Aristotle
suggested, by behaving like a virtuous person.
But, like automatic copying, deliberate selective
imitation does not always operate benignly. Selective imi-
tation can provide “Machiavellian” social advantages
(Byrne & Whiten 1988; Whiten & Byrne 1997). It can
steal not only instrumental successes but also cooperative
benefits from competitors. Suppose information about the
mental states of others is not transparently available.
A cooperative group can share certain behaviors by
which members identify one another, obtain cooperative
benefits, and exclude free-riding noncooperators. Coop-
erators may copy such identifying behaviors from other
cooperators.
Noncooperators could invade such a cooperative group
by selectively copying its identifying behaviors. They could
thus induce cooperation from group members while
failing to cooperate in return, deceptively obtaining coop-
erative benefits without paying the costs. Free riding via
deceptive copying partially appropriates cooperative
benefits based on in-group behavioral copying. While
greenbeard genes could produce genetically determined
analogues of such free riding (Dawkins 1982, p. 149),
selective copying provides the evolutionary advantages of
flexible free riding, which is not dependent on genes for
specific behaviors.
How can cooperative benefits be defended against free
riding through deceptive copying? An arms race between
behavioral signaling and deceptive copying in cooperative
games arguably produces pressure for imitative and mind-
reading abilities. As a result, certain solutions to coopera-
tive games, which require mindreading rather than mere
behavior prediction, may become available. Mindreading
can be based on behavioral evidence yet still have func-
tional advantages over behavior prediction (Hurley
2005a).
To elaborate: To counter invasion by increasingly soph-
isticated deceptive mimics, mutual recognition processes
among cooperators would move progressively further
from copying and detecting superficial behaviors and
toward more subtle and covert imitation and detection
of underlying mental causes of behavior. Mere behavior
reading would move toward ever-smarter reading of beha-
vioral evidence for intentions. Mere copying would in turn
become more creative and flexible, with means/ends
structure: imitation. This arms race could produce
capacities for mindreading and intersubjective identifi-
cation via covert mirroring, albeit based on subtle beha-
vioral perceptions (cf. Krebs & Dawkins 1984).
The advance from cooperation plus deceptive copying
via imitation to mindreading is significant for enabling
cooperation and obtaining its benefits. Certain solutions
to collective action problems effectively require recogniz-
ing and identify with others’ mental states. A simple
self-referential mirror heuristic8
for non-iterated Prison-
ers’ Dilemmas (PDs) says: cooperate only with any
others you meet who act on this same rule (Howard’s
mirror strategy [Howard 1988]; Danielson’s self-same
cooperation [Danielson 1991; 19929
). When another
player doesn’t share your mirror heuristic, you don’t
cooperate with him. Famously, Tit-for-Tat can outperform
Defection in iterated PDs, where given players meet
repeatedly; but mirror heuristics outperform Defection
even in non-iterated PDs, where given players don’t
meet again.10
Mirror heuristics effectively require mindreading: dis-
covering another player’s intention, not simply predicting
his behavior (Danielson 1992, pp. 75–82; Schmitt &
Grammar 1997).11
They are conditional metaheuristics:
they explicitly condition cooperation on the other’s operat-
ive heuristic itself, not on his predicted behavior. (Tit-for-
Tat requires not mindreading but memory of a given
Hurley: The shared circuits model
8 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
9. player’s behavior in past games.) Employing a mirror
heuristic requires discerning, more or less reliably,
whether others are operating on a mirror heuristic – a
general intention or rule of choice. Which choices mir-
rorers should make are not determined by predicting
what others will do; mirrorers need to know whether
others have the intentions of a mirrorer before they can
determine what to do. Participants in mirror-based
cooperation must not only be mindreaders, but also be
able to identify, more or less reliably, other mindreaders.
In non-iterated games, the mindreader has not pre-
viously played with the player she is mindreading, so
cannot refer to memories of their past play. In informa-
tionally clouded social environments, mindreading is
based on evidence from observing behavior, which may
be subtle and/or deceptive. But mindreading need not
be foolproof to provide mirror-based cooperative benefits
to individual mirrorers and groups of them; the benefits
would vary with the accuracy of mindreading (cf. Danielson
1992, pp. 157ff.).
What is the difference between genuine mindreading
and mere smart behavior reading? Many social problems
that animals face can be solved via behavior-circumstance
correlations and behavioral predictions, without postulat-
ing mediating mental states. What problem-solving press-
ures are addressed by additionally attributing mental
states to explain observed behavior? (Cf. Call & Tomasello
1999; Hare et al. 2000; 2001; Heyes 1998; Heyes &
Dickinson 1993; Hurley 2006b; Povinelli 1996; Povinelli
& Vonk 2006; Sterelny 2003, pp. 67ff.; Tomasello & Call
2006; Whiten 1996; 1997.)
Mental state attributions may support more flexible
behavior prediction in novel conditions. But mirror meta-
heuristics show that mindreading’s function in enabling
cooperation goes beyond providing better predictions of
behavior. As explained, mirror metaheuristics do not
require predicting other players’ behavior per se, but
rather, ascertaining the heuristic they use. Such mindread-
ing may be done by observing others’ behavior, but that
does not mean that its function is only behavior prediction,
or that it has the same functions as behavior prediction.
Mindreading can function to enable cooperation in a way
that merely predicting behavior cannot (Danielson 1992,
p. 82), even if mindreading is based on behavioral evi-
dence. This is why the emergence of mindreading, via imi-
tation, from an arms race between cooperative and
deceptive copying is significant for enabling cooperation.
2.3.4. Mindreading. What more can be said about the
possible functions of imitation in relation to mindreading?
Human mirror systems may be part of the mechanisms
for understanding observed actions and intersubjective
empathy. Observing another act primes your motor
system to copy, even if overt copying is inhibited. Covert
copying is a kind of process-driven simulation, which
uses offline the processes that would be used actually to
copy the observed action, but it inhibits motor output.
This direct resonance with another’s action provides a fun-
damental similarity between yourself and other agents that
enables the understanding of another’s actions as instru-
mentally structured. Mirror systems also provide a plaus-
ible neural basis for emotional empathy and
understanding (see Adolphs 2002; Decety & Chaminade
2003; 2005; Gallese 2001; 2005; Gallese & Goldman
1998; Goldman 2005; Gordon 1995a; 1995b; Iacoboni
2005; Iacoboni et al. 2005; Meltzoff 2005; Preston & de
Waal 2002; Rizzolatti 2005; Williams et al. 2001).
Within this broad perspective, I shall compare the views
of Gallese, Meltzoff, Gordon, and Tomasello on simulation
theory versus theory theory and on relations between imi-
tation and mindreading. And I will preview how SCM
reconciles opposed views on both topics. An outline
follows.
2.3.4.1. Simulation theory (ST) versus theory
theory (TT).
Gallese views mirror systems as enabling broad interperso-
nal empathy by implementing primitive intersubjective
information, prior to differentiation of self from other.
Meltzoff views early imitation as foundational for the
ability to understand other agents: In imitation my acts
are directly, noninferentially identified with others’
acts; I then associate my acts with my mental states
and infer a similar association in others.
TT accounts of mindreading invoke laws and inferences
about mental states and behavior, whereas on ST
accounts mindreaders use their own psychological pro-
cesses offline to attribute similar mental states or
actions to others.
Underived similarity between one’s own and others’ acts is
shared ground between Meltzoff’s TT account of mind-
reading based on early imitation and ST accounts based
on inhibited copying.
Gordon criticizes first-person- to third-person-inference
accounts of early imitation’s role in mindreading, for
taking the self/other distinction for granted.
In Gordon’s ST view of how mindreading involves offline
imitation, “constitutive mirroring” “multiplies the first
person” by reference to a shared scheme of reasons.
First reconciliation of ST and TT: Foundations of inter-
subjectivity and the self/other distinction can be pro-
vided by simulative mirroring (SCM’s layers 3 and 4),
although richer self/other and other/other distinctions
depend on interpretation, theorizing, and inference
(layer 5 and beyond).
2.3.4.2. Relations between imitation and action
understanding.
Tomasello and Carpenter’s view that imitation depends on
action understanding contrasts with views of action
understanding as depending on imitation.
Second reconciliation, concerning relations between imi-
tation and action understanding: Simple mirroring, of
goals or movements, can express a fundamental inter-
subjectivity, enabling simple forms of action under-
standing and providing elements of more complex
imitative mirroring with flexible instrumental structure,
which in turn contributes to more articulated, instru-
mentally structured understanding of other agents and
their minds (layers 3 and 4).
2.3.5. Simulation theory (ST) versus theory theory
(TT). On Gallese’s (2001; 2005) shared manifold hypoth-
esis, mirror systems enable various aspects of interperso-
nal understanding and empathy. Mirror systems develop
from the way biological control systems model interactions
between organisms and their environments. They provide
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 9
10. the neural basis of a primitive intersubjective information
space or “shared manifold” that is prior to a self/other dis-
tinction both phylogenetically and ontogenetically, yet
preserved in human adults (SCM’s layer 3 incorporates
this feature). The shared manifold underwrites automatic
intersubjective identifications across different perceptual
modalities and action, but also for sensations and
emotions: There is evidence of mirror mechanisms for
pain and disgust, and hearing anger expressed increases
the activation of muscles used to express anger.
Empathy involves a common scheme of reasons under
which persons, self and others, are intelligible, rather
than recognition that others’ bodies also have minds.
Meltzoff (2005) argues that early imitation and its
enabling mechanisms begot understanding of other
agents, not vice versa. In his view, ability to understand
other minds has innate foundations, but develops in
stages, in which imitation plays a critical role. Infants
have a primitive ability to recognize being imitated and
to imitate, and hence to distinguish people from other
things and recognize equivalences between acts of self
and other. This initial bridge between self and other pro-
vides privileged access to people that we don’t have in
relation to other things; it develops in three stages. First,
own acts are linked to others’ similar acts supramodally,
as evidenced by newborns’ imitation of others’ facial ges-
tures. Second, own acts of certain kinds are linked bidirec-
tionally to own mental states of certain kinds, through
learning. Third, others’ similar acts are linked to others’
similar mental states. This early process is conceived not
as formal reasoning, but as processing the other as “like
me.” It gets mindreading started on understanding
agency and closely associated mental states: for example,
intentions, emotions, and desires. Meltzoff emphasizes
that mindreading isn’t all or nothing. (Tomasello [1999]
makes similar claims for nonhuman animals.) Understand-
ing mental states further from action, like false beliefs,
comes in later development.
Meltzoff’s view of mindreading is usually put in the
theory theory (TT) rather than simulation theory (ST)
category, but it has ST as well as TT aspects. TT regards
commonsense psychology as a proto-scientific theory. It
represents knowledge as laws about mental states and
behavior that can be known innately or discovered by
testing hypotheses against evidence. Specific mental
states and behaviors are inferred from other mental
states and behaviors by means of such laws; the process
does not depend on copying. On ST, mindreading starts
with taking another’s perspective and generating
“pretend” mental states or behavior that match the
other’s. These offline states are not objects of theoretical
inference. Rather, they are entered into the simulator’s
own psychological and decision-making processes, which
are held offline to produce further simulated mental
states and behavior that are then assigned to the other.
Further behavior by the other can be predicted, or
mental states attributed that explain his observed beha-
vior. Such simulation is an extension of practical abilities
rather than a theoretical exercise: it copies the other’s
states and uses the copies in the simulator’s decision-
making equipment, instead of using laws to infer the
other’s states (Davies & Stone 1995a; 1995b).
Meltzoff’s three-stage process can be restated in expli-
citly TT terms. First, innate equivalence between my
own and others’ acts (exploited by early imitation and rec-
ognition of being imitated) provides a fundamental, under-
ived similarity between some acts (by myself) and other
acts (by another). Second, first-person experience provides
laws linking one’s own acts and own mental states. Third, it
is inferred that another’s acts and mental states are law-
fully linked in the same ways as my similar acts and
mental states are linked. Proceeding through stages 2
and 3, we find inferences from first-person mind-behavior
links to similar third-person links as in traditional
arguments from analogy.12
“The crux of the ‘like-me
hypothesis’ is that infants may use their own intentional
actions as a framework for interpreting the intentional
actions of others” (Meltzoff 2005, p. 75). For example,
12-month-old infants follow a model’s “gaze” significantly
less when the model’s eyes are closed rather than open,
but only similarly refrain from following the “gaze” of a
blindfolded model after they are given first-person experi-
ence with blindfolds.
However, as Meltzoff points out (personal communi-
cation), there is no first-person to third-person inference
at stage 1. The initial bidirectional self-other linkage,
expressed in early imitation and recognition of being imi-
tated, is via a supramodal common code for observed and
observer’s acts that’s direct and noninferential (Meltzoff &
Moore 1997). Stage 1 of Meltzoff’s view thus has important
common ground with ST: In covert offline copying, direct
noninferential resonance with another’s action with inhib-
ited motor output enables understanding of the other’s
action. But such direct noninferential resonance can also
occur in overt copying, as Meltzoff postulates; copying
can provide information for understanding another’s
actions, even when not inhibited and serving other func-
tions. Mindreading’s foundation at Meltzoff’s first stage
is noninferentially direct, not theoretically derived. His
view shares this nontheoretical basis, at its first, online
copying stage, with ST views of mindreading as based on
offline copying, though they diverge on how mindreading
develops further. If mindreading develops in stages, theor-
etical inference can enter later, increasing with
development.
While Meltzoff’s theory theory involves first-person to
third-person inference, Gordon’s “radical” simulation
theory (see Gordon 1986; 1995; 1995a; 1996; 2002;
2005) explicitly rejects it, and provides a different view
of the relations between imitation and mindreading. In
constitutive mirroring, a copied motor pattern is part of
the perception of another’s action, though overt move-
ment may be inhibited. Gordon finds constitutive mirror-
ing in Gallese’s primitive intersubjective “we”-space, the
basis of empathy that implicitly expresses similarity of
self and other rather than their distinctness. When consti-
tutive mirroring imposes first-person phenomena,
a process of analysis by synthesis occurs whereby another’s
observed behavior and the self’s matching response – part
of the very perception of the other’s behavior – become
intelligible together, in the same process. When I see
you reach to pick up the ringing phone, your act and my
matching response are made sense of together within a
scheme of reasons that is fundamentally common to
persons. As Gordon (2005) puts it, I don’t infer from the
first to the third person, but rather multiply the first
person. To understand what I or another believes, per-
ceives, or intends, I look out at the world and the
Hurley: The shared circuits model
10 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
11. reasons it provides, though in the case of others I imagina-
tively recenter to the other’s perspective (Gordon 1995a).
Gordon criticizes first-person to third-person inference
in Meltzoff’s account not because it attributes similarity to
one’s own and others’ acts or experiences, but because it
requires that they be identified and distinguished. In
Meltzoff’s stage 1, there is innate equivalence between
acts of self and other; this stage may involve constitutive
mirroring, as in Gallese’s primitive shared manifold. But
later stages of Meltzoff’s account, where analogical infer-
ence occurs, require that self and other also be distin-
guished: If this kind of act by me is linked to my mental
states of a certain kind, then a similar (as per stage 1)
kind of act by another is also linked to her mental states
of a similar kind. Gordon explains that I cannot infer ana-
logically from a to b unless I can distinguish a from b. He is
skeptical that infants have this capacity, though mature
imitative mirroring may involve such inference (Gordon
2005).
Pure ST views of mindreading are standardly criticized
for lacking resources to explain how mature mindreaders
distinguish and identify people and track which actions
and mental states are whose. Gordon suggests that mul-
tiple first persons are distinguished and tracked in the per-
sonal-level process of making others intelligible, avoiding
incoherence under the common scheme of reasons (see
also Hurley 1998, part 1; 1989). Mental states that don’t
make sense together are assigned to different persons.
Can this be done in pure simulation mode, without theo-
rizing? Simulation supposedly uses practical abilities
rather than theorizing about actions. How does interpret-
ing an action to make sense of it differ from theorizing
about it? When I use practical reason offline in interpret-
ative mindreading, I don’t formulate normative laws from
which I make inferences; rather, I activate my normative
and deliberative dispositions. As Millikan might say
(Millikan 2005), my thought about another’s action isn’t
wholly separate from my entertaining that action.
SCM will suggest a reconciliation between ST and TT.
The fundamental similarity between self and other is
understood in terms not of theorizing but mirroring
(as in Gallese’s shared manifold, Gordon’s constitutive
mirroring, Meltzoff’s innate self-other equivalence, and
SCM’s layer 3). Such primitive intersubjectivity persists
into adulthood, providing a basis for mature empathy
and mindreading, as Gallese holds. The informational
origin of the self/other distinction is understood in terms
of monitoring whether mirroring is inhibited (layer 4).
As mindreading develops, it also employs a richer self/
other distinction, as when children come to distinguish
imitating from being imitated (see Decety & Chaminade
2005), or to attribute beliefs different from their own to
others. Mature personal-level mindreading requires abil-
ities to distinguish, identify, and track multiple other
persons, to assign acts and mental states to them in an
interpretative process, and to entertain multiple possible
acts by multiple other persons (layer 5). If decentering
from me-here-now creates a trail to others and other poss-
ible actions, mature mindreading creates multiple branch-
ing and interacting trails. Negotiating these, by using the
full range of distinctions and identifications required by
mature mindreading, probably demands theoretical
resources, even though the subpersonal enabling foun-
dations of intersubjectivity are found in mirroring, and
of the self/other distinction in monitored simulation.
SCM explains how mirroring and simulation can provide
foundations for mindreading on which theorizing builds.
2.3.6. Relations between imitation and action
understanding. How are imitation and action understand-
ing related to each other? On imitation-first views, imita-
tion underwrites early mindreading abilities. Gallese,
Meltzoff, Gordon, and Goldman stress the contribution
of imitation to understanding other agents. By contrast,
understanding-first views emphasize the way imitative
learning depends on action understanding and intention
reading (Carpenter et al. 1998; Rizzolatti 2005; Tomasello
& Carpenter 2005). Recent paradigms with children
where the demonstrated action is unsuccessful or acciden-
tal (Meltzoff 1995) distinguish imitation from other forms
of social learning more clearly than the two-action method
does. If the observer copies what was intended even
though it wasn’t achieved, as opposed to copying only
the observed movements or the observed though unin-
tended result, that suggests the observer understands the
intentional structure of the observed action. Tomasello
and Carpenter argue that intention reading is needed to
explain what is copied by imitators when the modeled
behavior is the same across conditions while the
modeled intention varies. In their view, results from
various paradigms are most parsimoniously explained by
holding that children use their understanding of intentions
to imitate.
Imitation-first and understanding-first views are not
necessarily opposed; each may tell only part of the story.
SCM provides a framework for their reconciliation,
accommodating both views at different points in its
layered architecture. Different types of copying, and
covert forms of each that enable corresponding types of
understanding, can dovetail over evolution and develop-
ment, building on one another reciprocally, with increas-
ing instrumental structure in both action and
understanding over successive stages. A simpler form of
copying can precede a simpler form of understanding,
which precedes a more complex form of copying, which
precedes a more complex form of understanding (the
ordering can be interpreted phylogenetically or ontogen-
etically). Start, say, with goal mirroring and emulation.
Covert goal mirroring can then enable understanding
the goals of observed action. Such goal understanding,
along with mirroring of movements, may be needed for
instrumentally articulated imitation (understanding first).
But richer instrumental understanding, of how observed
means contribute to observed ends, may involve covert
imitation (imitation first). In SCM, self-other similarities
expressed by mirroring, whether more or less structured,
are informationally prior to the self/other distinction
required for understanding action as another’s.
3. Part 2. The shared circuits model
The shared circuits model (SCM) shows how subpersonal
resources for control, mirroring, and simulation can
enable the distinctively human sociocognitive skills of imi-
tation, deliberation, and mindreading. The model has
intertwined empirical and philosophical aims. One aim is
to provide a unified framework for the various strands of
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 11
12. empirical evidence and theorizing surveyed thus far.
Another is to illustrate the philosophical view that embo-
died cognition can emerge from active perception, avoid-
ing the “classical sandwich” architecture, which insulates
central cognition from the world between twin buffers of
perceptual input and behavioral output (Hurley 1998;
2001). It does this, recall, by addressing a higher-order
theoretical question, about how it is possible for subperso-
nal processes to enable certain personal-level abilities: in
particular, how it is possible to build subpersonal
resources for sociocognitive skills on those for active
perception.
SCM thus provides a generic heuristic framework for
specific first-order hypotheses, about how particular socio-
cognitive capacities map onto specific layers of the model
or develop in phylogenetic or ontogenetic time. SCM itself
does not articulate specific first-order hypotheses, but
does make some general predictions: for example, of
neural systems for mirroring based on those for instru-
mental prediction, and of the priority of online over
offline mirroring. Nor is SCM exclusive; important work
in enabling persons’ cognitive capacities is done by other
processes, including linguistic processes. The point is to
illustrate how it is possible for important cognitive
resources to emerge from active perception. Although
SCM is therefore somewhat abstract, in accord with its
higher-order theoretical aim, I suggest in this target
article how it lends itself to more specific empirical predic-
tions, and hope the commentaries will offer suggestions as
well. Details follow, layer by layer.
3.1. Layer 1: Basic adaptive feedback control
SCM begins with specific comparator feedback control
systems. A comparator system generates outputs that are
means to a target, by establishing an instrumental associ-
ation between outputs and their results. For example,
a thermostat compares a target signal with an input
signal. If they don’t match, system output is adjusted and
the resulting change in input signal or feedback is
tracked. Input continues to be recompared with target
and output readjusted, to minimize mismatch to target.
The elements of such control are (Fig. 1):
1. A target or reference signal (e.g., target room temp-
erature for a thermostat).
2. An input signal (e.g., actual room temperature), the
joint result of elements 3 and 5.
3. Exogenous environmental events (e.g., nightfall).
4. A comparator, which determines whether target and
input signals match and the direction and degree of any
mismatch or error (e.g., the room is still five degrees
below target temperature).
5. The output of the control system (e.g., the level of
heat output), regulated by comparison between target
and input signals (e.g., heat output is increased if
measured room temperature is below target).
6. A feedback loop by which output has effects on suc-
ceeding input signals (e.g., measured room temperature
rises when heat output increases).
Feedback control is adaptive; output is adjusted to com-
pensate for changing exogenous influences, keeping
sensed input close to target. Under different exogenous
influences, feedback calls for differing outputs to achieve
the target; when the weather changes, a thermostat
adjusts heat output to maintain the target temperature.
Feedback at layer 1 operates in real space and time, and
therefore can be slow (e.g., a room takes time to warm
up after the heat is turned up). A control system
implements a mapping from the target, in the context of
actual input, to output, thus specifying the means for
approaching the target in given circumstances. Inverse
model is engineering terminology for this instrumental
mapping.
Net sensed input results from the system’s output plus
independent environmental influences. In organisms,
reafferent feedback carries input resulting from the organ-
ism’s own activity, whereas exafferent input results from
exogenous events. Reafference includes visual and pro-
prioceptive inputs resulting from movements of one’s
hands, movement through space, manipulation of
objects, and so on. Exafference includes visual inputs
resulting from environmental events, such as movements
by others in a social group. However, at layer 1, infor-
mation distinguishing reafference from exafference is not
available.
Feedback control is a cyclical and dynamic process, with
no nonarbitrary start, finish, or discrete steps; input is as
much an effect as a cause of output (Marken 2002;
Powers 1973). Control depends on dynamic relations
among inputs and outputs. Information about inputs is
not segregated from information about outputs; this
blending of information is preserved and extended in the
informational dynamics of further layers. Perception and
action arise from and share this fundamental informational
dynamics (Hurley 1998; 2001).
Specific means/ends associations or instrumental map-
pings can be chained (output A is the means to controlled
result B, while B in turn is the means to controlled result
C, and so on) or organized into hierarchies. There are
independently determined evolutionary, developmental,
and individual differences in the grain and complexity
of the possible control sequences and hierarchies of
different creatures.
3.2. Layer 2: Simulative prediction of effects for
improved control
Real-time feedback can be slow and produce overshoot-
ing. Control functions can be speeded and smoothed by
adding simulative predictions to a comparator system
(Grush 2004; Miall 2003): Instrumental output-result
associations can then be activated predictively, simulating
the effects of specific outputs for informational purposes.
Figure 1. Layer 1: Basic adaptive feedback control.
Hurley: The shared circuits model
12 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
13. Over time, associations are established between output
and subsequent input in certain contexts, so that copies
of motor-output signals can evoke associated input
signals. An inner loop maps copies of output signals
directly onto “expected” input signals – means to results
(Fig. 2; new aspects italicized). Forward model is engin-
eering terminology for this mapping; copies of output
signals in organisms are called efference copy. This subper-
sonal process simulates feedback – predicts the results of
output on input. Prediction can occur during actual action,
to smooth a behavioral trajectory by anticipating feedback,
or prior to action, to provide information about alternative
possible actions (layer 4).
A general improvement in instrumental control results.
A control system with predictive simulation no longer
need await actual feedback. A thermostat can turn the
heat down before the target is reached, avoiding over-
shooting; hand movement can be initiated in accord with
predictions of retinal signals based on eye movements.
When real and simulated results don’t match, a local
switch can default back to actual feedback control while
predictive simulations are fine-tuned to improve sub-
sequent predictions.13
For purposes of online instrumen-
tal control, the system need not monitor continuously or
access globally whether it is using actual or simulated
feedback.
Comparisons can now be made not just between targets
and actual results of action, but between the latter
and anticipated results. This permits reafference to be
distinguished from exafference: Information about the
organism’s goal-directed behavior can be distinguished
from information about environmental events. Consider
the familiar ambiguity: When my train pulls out of the
station, I register movement relative to the train on the
next platform, but this does not itself provide information
about whether my train or the neighboring train is moving.
Comparison of predicted feedback from action (efference
copy) with actual feedback provides resources to resolve
an analogous subpersonal ambiguity (between reafference
and exafference), and hence to distinguish the self’s
activity from environmental events.14
This subpersonal
information can contribute to enabling the personal-level
distinction between action of the self and perception of
the world. The perception/action distinction emerges
from subpersonal informational dynamics between
world-to-animal inputs and animal-to-world outputs.
However, perception and action don’t map, respectively,
onto input and output (Hurley 1998). Rather, layer 2
inherits unsegregated information about inputs and
outputs from layer 1, and uses this blended information
in enabling the perception/action distinction. Perception
and action share these basic informational dynamics and
processing resources. SCM thus provides a dynamic
process version of a common coding view of perception
and action (sect. 2.2).
However, the system does not yet provide information
about similarities between the agent’s actions and actions
by others, nor information distinguishing the agent’s
actions from similar actions by other agents (as opposed
to distinguishing the agent’s actions from environmental
events in general). This suggests that a basic distinction
between action by self and perception of world, associated
with instrumental control functions, can be available to
creatures still lacking intersubjective information, a self/
other distinction, or mindreading abilities. There are
more and less fundamental layers of information about
self (self in SCM is neutral between persons and other
animals).
Some general predictions derive from layer 2. First,
neural mechanisms that implement sensorimotor affor-
dance associations (such as canonical neurons [sect. 2.2])
are predicted. Suppose an animal typically acts in ways
afforded by certain kinds of object: for example, eating
a particular food in a specific way. Copies of motor
signals for eating movements will be associated with a mul-
timodal class of exafferent and reafferent inputs deriving
from such objects and the agent’s eating of them. Cells
mediating this sensorimotor association could thus have
both sensory and motor fields and carry information
about objects’ affordances. Second, deficits in predictive
simulation functions should be associated with deficits
in distinguishing self from world and action from percep-
tion.15
A specific first-order prediction relating to layer 2
might be that capacities requiring information that dis-
tinguishes self from world are phylogenetically prior to
capacities for social learning and action understanding.
3.3. Layer 3: Mirroring for priming, emulation, and
imitation
SCM next postulates that instrumental output-result associ-
ations can be activated bilaterally, from effect to cause, as
well as from cause to effect. Not only do copies of motor
signals predictively simulate input signals (layer 2), but
input signals can evoke corresponding motor signals.16
To put it more technically, not only does efference copy
produce simulated reafferent input in forward models,
but input signals can evoke mirroring efference or motor
output. Mirroring in effect runs the predictive simulations
of forward models in reverse (Fig. 3).
Observed actions are thus mirrored in the observer; if
mirroring is sufficiently strong and not inhibited, overt
copying results. (Mirroring is here a functional subperso-
nal rather than neural description of behavioral priming
produced by observing action, and may be implemented
in neural mirror systems.) Mirroring within specific
control structures of differing grain and structure would
enable different copying capacities, observed across
various social species: mirroring of basic movements in
priming (Rizzolatti’s “low-level resonance”), mirroring of
goal-directed action or emulation (Rizzolatti’s “high-level
resonance”), and even full-fledged imitation (if mirrored
Figure 2. Layer 2: Simulative prediction of effects for improved
control.
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 13
14. elements of control structures are sufficiently articulated
and flexibly linked to provide the information needed for
social learning of novel means to a goal [see further on
in this section]).
Note the intimate relationship between the sharing of
circuits for self and other and for action and perception:
Layer 3’s shared informational dynamics for intersubjec-
tivity presupposes layer 2’s shared informational dynamics
for perception and action, which builds on layer 1’s
generic informational dynamics for sensorimotor control.
SCM explicitly builds shared resources for self and other
on those for action and perception. It thus integrates
W. Prinz’s shared information for perception and action
(though in terms of functional dynamics rather than
coding) with Gallese’s primitive intersubjective “we”-
centric information. By bringing together information
about motor causes of one’s own and others’ similar
observed actions, mirroring enables simulation of
means/ends associations from either direction: Observed
action retrodicts motor activation in the observer via mir-
roring of causes, which are associated with further results
via simulative prediction of effects. But although mirroring
makes information about action’s instrumental structure
accessible bilaterally (from acting and from observing
others act), mirroring does not yet distinguish own
action from observed action. At layer 3, self and other
share informational resources: intersubjective information
is subpersonally prior to the self/other distinction
(as Gallese holds [sect. 2.3]).
Instrumental mapping and mirroring both map input to
output; SCM distinguishes them functionally (neural
implementations may overlap). Instrumental mappings
have control functions: given certain inputs, they select
motor outputs that in turn produce inputs matching a
target. Although mirroring exploits instrumental control
structures and also produces motor outputs, given
certain inputs, it does not itself select outputs as means
to inputs that match observed action – or any other
target (cf. Peterson & Trapold 1982). Rather, SCM postu-
lates, mirroring is a by-product (via reversal) of predictive
simulations, which do have instrumental control functions.
However, the resulting automatic copying tendency has
evolutionary functions, and copying can be exapted for
cognitive functions associated with imitation, action
understanding, or signaling; these in turn can enable
advanced social (“Machiavellian”) forms of instrumental
control (see sects. 2.1 and 2.3). Whether specific mirroring
capacities are adaptive depends on their potential
functions for different social species under different evol-
utionary pressures.
How could mirroring arise? Consider first movements
that produce visual reafference. When creature A sees
her own hand movements, associations form between
copies of motor signals for these movements and visual
reafference from these movements. Cells mediating this
association can acquire congruent sensory and motor
fields. These cells would also fire if creature A receives
similar visual inputs from creature B’s similar hand move-
ments; the cells wouldn’t distinguish observer’s action
from observed action producing similar inputs. So, like
mirror neurons, they would fire both when A acts and
when she observes such similar action by B. They
mediate associations between copies of motor signals
and a class of inputs including both characteristic reaffer-
ence from the agent’s movement and similar exafference
from others’ similar movements.
How could mirroring arise for movements not seen by
their agents? This requires a correspondence between
one’s own and others’ similar acts, without reafferent feed-
back from own acts in the same modality as observations of
others’ acts. Facial movements normally produce proprio-
ceptive rather than visual feedback: How can one’s own
facial movements be associated with visual information
about similar observed facial movements, to enable
copying of facial expressions?
Several answers are possible (sect. 2.2), all compatible
with SCM. Some supramodal correspondences may be
innate (as in newborn copying; Meltzoff 2005; Meltzoff &
Moore 1997). Some may be acquired through experience
with mirrors, or with being imitated (Heyes 2005). Some
could be established via stimulus enhancement, as
follows. Suppose a creature repeatedly sees conspecifics
act a certain way; their actions draw its attention to the
typical objects of their actions, which evoke in the observer
an innate or previously acquired response. An indirect
association results between visual observations of others’
actions and one’s own similar action. This isn’t copying
initially; the object independently evokes others’ and
own similar acts. But the indirect, object-mediated associ-
ation between own and others’ similar acts may become
direct with repetition. Cells mediating it could thus
acquire similar sensory and motor fields, so that observing
another’s act primes similar action by the observer. There-
fore, stimulus enhancement could develop into copying,
and an indirect link via an enhanced stimulus into a
direct sensorimotor mirroring link.17
SCM is neutral
about whether such correspondences are innate, acquired,
or both.
SCM does not describe one all-inclusive structure, but
has multiple instances for specific movements and
results, at various points along different means/ends
chains (cf. Fogassi et al. 2005; Wolpert et al. 2003). The
ends/means distinction is relative and applies along
spectra of means/ends links in which basic movements
are means to proximal results that are means to more
distal results. Such control spectra can vary in grain.
SCM could apply at successive points along a control spec-
trum, or between spectra (with the right neural connec-
tions); one circuit’s target could be the means to next
circuit’s target. A network of control spectra could
support hierarchical control and flexible recombination
Figure 3. Layer 3: Mirroring for priming, emulation, and
imitation.
Hurley: The shared circuits model
14 BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1
15. of means and ends. With mirroring added, it could enable
movement priming, emulation, or imitation.
SCM predicts that capacities for specific kinds of
copying and social learning will vary across species and
development with
(A) the grain and complexity of instrumental control
capacities, and
(B) which of these have associated mirroring functions and
how richly and flexibly they are linked (see discussion
following; Csibra 2005).
How do these two influences operate?
(A) Different animals are equipped to different degrees
with capacities for instrumental control and associated
predictive simulation. This variation reflects potential
means/ends chains of differing grains and lengths and
with differing degrees of lateral connectivity to enable
novel combination of means and ends as instrumentally
appropriate. Animals suitably equipped with control func-
tions by evolution and development could form chains of
simulative predictions, resulting in information such as:
this tail movement has that effect on weight over legs,
facilitating this movement trajectory in relation to that
gazelle, and so on – with eating the prey later in the
chain. A similar chain could lead from knee movement
to winning a slalom race.
Mirroring operates such instrumental associations in
reverse. Mirroring the cause of another’s movement, or
resulting relationship to an object, could enable movement
priming, goal emulation, or even full-fledged imitation
(if instrumental control and associated mirroring functions
are sufficiently articulated and flexible). Combined with
further information distinguishing self from other (layer 4),
simulative mirroring can provide information to enable
understanding of others’ observed movements as instru-
mental actions with intentional, means/end structure.
Simpler control and predictive simulation capacities, with
shorter, coarser, means/ends chains, limit correspondingly
an animal’s potential for mirroring and related functions.
Whether observed movement primes copying or is
recognized as goal-directed depends on, inter alia, the
instrumental capacity potentially available for mirroring
and simulative functions. Mirroring and simulation might
provide information about the goals of certain observed
movements, given fine-grained, complex means/ends
associations, but not given coarser control capacities. No
doubt a monkey can move her hand to grasp a piece of
sushi and move it to her mouth to eat it. But I can move
my hand to operate chopsticks to pick up sushi to dip it
in soy sauce and then move it to my mouth to eat it, in
order to impress my boss; given associated simulative mir-
roring functions, I may start to resent you for eating the
last piece of sushi as soon as you reach for your chopsticks.
(B) While mirroring of instrumental associations can
provide information about the instrumental structure of
observed movements, mirroring does not itself determine
which instrumental associations, or predictive simulations
thereof, are in place and hence potentially exploitable by
mirroring. Control processes can be neurally distributed,
with components of an articulated mean/ends chain pro-
cessed in different brain areas – control of fine move-
ments versus gaze versus posture versus whole body
movement versus external objects. Some such neural
areas may have mirroring as well as predictive simulation
capacities, whereas others do not. Whether mirroring is
associated with particular control structures will vary
across species and development, as a result of evolutionary
and developmental processes, yielding different capacities
for copying and for generating information about goals of
observed actions.
We can understand the enabling of movement priming,
goal emulation, or imitation in terms of mirroring that
exploits different control structures. Mirroring associated
with more basic means/ends links (Rizzolatti’s ‘low-level
resonance’) predicts priming of basic movements; mirror-
ing associated with less basic means/ends links predicts
priming of less basic movements or results of more basic
movements. Finger movements are means to chopstick
deployment; when I can control chopsticks, chopstick
deployment is a means to sushi eating, which could be a
means to some further social result. Suppose I eat sushi
by deploying chopsticks by moving my fingers a certain
way. You watch. If seeing me move my fingers generates
mirroring motor activation, then you are primed to move
your fingers similarly. Such movement priming predicts
interference if you watch me doing X while you are
doing Y.
If you can already control chopsticks, less basic mirror-
ing and prime chopstick deployment can occur. Such
priming could be goal-mediated: Your chopstick deploy-
ment could mirror mine when sushi-eating results (even
if no sushi is visible), but not when the results are unre-
lated to sushi (cf. Umiltà et al. 2001). Goal emulation
could be enabled by mirroring midway along a control
spectrum of proximal results of movements (rather than
basic movements themselves) that are in turn means to
more distal results (cf. Rizzolatti’s “high-level resonance”).
Such midway mirroring would generate motor activation
associated with a corresponding midway goal for the
observer (rather than with similar basic movements in
the observer). As well as enabling emulation, this would
contribute information for understanding observed
action as goal directed (layer 4).
The phylogenetically rare capacity for imitative learning
requires flexible means/ends associations; priming and
emulation, respectively, provide its ends and means com-
ponents (sect. 2.1). Articulation within and linkages
between mirror circuits determine whether mirroring
enables imitative learning. An animal with mirror circuits
along a chain of means/ends associations may never
have used a certain means to a given goal. But if an
observed novel means to that goal is mirrored, and
neural links permit specific mirroring activations to be
flexibly combined with targets, that goal and those mir-
rored means may be newly associated, capturing infor-
mation about novel instrumental structure in observed
action and enabling imitative learning. You might learn
to use chopsticks by watching me; emulation plus trial
and error would be bettered by mirroring that primes
your finger movements towards those you see me make
that are associated with the target of chopstick
deployment.
Children’s greater imitative tendencies, compared with
chimps, may depend not on the presence versus absence
of a mirror system, but on articulated relationships
among multiple mirror circuits, permitting recombinant
structure in social learning (sect. 2.1; cf. Barkley [2001]
on recombinant structure in executive functions and
their relation to imitation). SCM predicts correctly that
Hurley: The shared circuits model
BEHAVIORAL AND BRAIN SCIENCES (2008) 31:1 15